Research On Single Haze-image Restoration With Multiple Priors

The computer vision technology has been applied widely and deeply in every field of daily life, but the computer vision collection equipment is affected easily by the haze, so the quality of the collected images have been supressed. For specific performance, the contrast decreased, the color offset, etc. So the extended applications have been influence greatly. Therefore, studying the haze-image restoration algorithm and removing the influence of the haze environment to the computer vision collecting device, have great significance to engineering application and scientific research.There are a great many of single haze-image recovery algorithms, and they are mainly divided into the contrast enhancement algorithm based on subjective vision and the inverse problem algorithm based on the atmospheric imaging model. There are some disadvantages of dark brightness and fuzzy edge details in the recovery results for the traditional algorithms, and the subjective visual effect is not perfect. So, it does not meet the requirements of the practical engineering applications. To solve these above problems, from the forward atmospheric imaging process, multiple priors are introduced to improve the brightness andkeep the edge details of the recovered image. The main contents and achievements of this paper as follow:1. The forward atmospheric imaging process is studied deeply.And the imaging model where the illumination usually is considered to be constant is generalized to suit for the condition of the illumination with non-constant value. Afterreviewing systematically the mainstream algorithm of the contrast enhancement based on subjective vision and the inverse problem algorithm based on the atmospheric imaging model, and analyzing and referring the priories in the above algorithm, a new objective function of single haze-image recovery is structured.2. The brightness of these recovery results of these traditional single haze-image recovery algorithm is low, so the visual effects and the following object extraction are influenced. For this, the airlight added image is obtained with the dark channel prior, then the airlight attenuated of every pixel can be estimated adaptively with the bright channel prior, and the airlight is corrected with priories. At last, the objective function is solved with the least square method. The haze can be wiped out, and at the same time, the brightness can be improved adaptively.3. The details will be fuzzy with the12norm constraint of the objects reflectivity, and it is difficult to separate subjects and extract margin. Therefore, the sparse prior of object edge is applied, and the l2norm restraint of the objects reflectivity is replaced with l0norm to maintain the edge structure information of images, and the corresponding iterative threshold values algorithm is provided. The experiments show that the structural information such as image edge can be remained preferably.